Aligning the behavior of Large language models (LLMs) with human intentions and values remains a critical challenge. Reinforcement learning from human feedback (RLHF) aligns LLMs by training a reward model (RM) on human preferences and fine-tuning the LLMs to maximize RM feedback. Despite its effectiveness and popularity, RLHF is prone to biased local optimization. It means RM fails to provide feedback that accurately aligns with human preference, causing LLMs to explore unexpected generalizations, and failing to achieve alignment objectives. To mitigate this issue, we propose a novel \textit{sequence-to-sequence (seq2seq) reward modeling} method. Its key insight is that learning from language feedback rather than scalar feedback improves RLHF without additional annotations. We replaced the reward modeling target from binary maximum likelihood estimation (MLE) with sequence MLE. This method enables richer and fine-grained language feedback without additional annotations, models, or training stages. Our experiments demonstrated its effectiveness, specifically, reducing the refusal-to-response paradigm in single-turn safety dialogues and the long-response bias in text summarization tasks. We provide further analysis that seq2seq RM improves RLHF performance across 2B and 7B LLMs on 3 NLP tasks, achieving an average win rate of 76.9\%. We further show that seq2seq RM can still improve the performance of RLHF under out-of-distribution prompts.

Data valuation seeks to answer the important question, "How much is this data worth?" Existing data valuation methods have largely focused on discriminative models, primarily examining data value through the lens of its utility in training. However, with the push for ever-larger language models, relying on valuation methods that require training becomes increasingly expensive and dependent on specific techniques. We propose an alternative perspective on the data value problem for language models, centering around the plausibility of the data. We posit that data holds lesser value if it can be plausibly generated by the model itself. Starting from some intuitive criteria that align with our notions of valuable data, we develop a novel value function that is computationally tractable and derived from first principles with provable properties. We conduct a theoretical analysis of our value function and evaluate it across multiple scenarios and datasets.

The dominant practice of AI alignment assumes (1) that preferences are an adequate representation of human values, (2) that human rationality can be understood in terms of maximizing the satisfaction of preferences, and (3) that AI systems should be aligned with the preferences of one or more humans to ensure that they behave safely and in accordance with our values. Whether implicitly followed or explicitly endorsed, these commitments constitute what we term a preferentist approach to AI alignment. In this paper, we characterize and challenge the preferentist approach, describing conceptual and technical alternatives that are ripe for further research. We first survey the limits of rational choice theory as a descriptive model, explaining how preferences fail to capture the thick semantic content of human values, and how utility representations neglect the possible incommensurability of those values. We then critique the normativity of expected utility theory (EUT) for humans and AI, drawing upon arguments showing how rational agents need not comply with EUT, while highlighting how EUT is silent on which preferences are normatively acceptable. Finally, we argue that these limitations motivate a reframing of the targets of AI alignment: Instead of alignment with the preferences of a human user, developer, or humanity-writ-large, AI systems should be aligned with normative standards appropriate to their social roles, such as the role of a general-purpose assistant. Furthermore, these standards should be negotiated and agreed upon by all relevant stakeholders. On this alternative conception of alignment, a multiplicity of AI systems will be able to serve diverse ends, aligned with normative standards that promote mutual benefit and limit harm despite our plural and divergent values.

This paper analyzes reinforcement learning (RL) algorithms for Markov decision processes (MDPs) under the average-reward criterion. We focus on Q-learning algorithms based on relative value iteration (RVI), which are model-free stochastic analogues of the classical RVI method for average-reward MDPs. These algorithms have low per-iteration complexity, making them well-suited for large state space problems. We extend the almost-sure convergence analysis of RVI Q-learning algorithms developed by Abounadi, Bertsekas, and Borkar (2001) from unichain to weakly communicating MDPs. This extension is important both practically and theoretically: weakly communicating MDPs cover a much broader range of applications compared to unichain MDPs, and their optimality equations have a richer solution structure (with multiple degrees of freedom), introducing additional complexity in proving algorithmic convergence. We also characterize the sets to which RVI Q-learning algorithms converge, showing that they are compact, connected, potentially nonconvex, and comprised of solutions to the average-reward optimality equation, with exactly one less degree of freedom than the general solution set of this equation. Furthermore, we extend our analysis to two RVI-based hierarchical average-reward RL algorithms using the options framework, proving their almost-sure convergence and characterizing their sets of convergence under the assumption that the underlying semi-Markov decision process is weakly communicating.

Balancing computational efficiency with robust predictive performance is crucial in supervised learning, especially for critical applications. Standard deep learning models, while accurate and scalable, often lack probabilistic features like calibrated predictions and uncertainty quantification. Bayesian methods address these issues but can be computationally expensive as model and data complexity increase. Previous work shows that fast variational methods can reduce the compute requirements of Bayesian methods by eliminating the need for gradient computation or sampling, but are often limited to simple models. We demonstrate that conditional mixture networks (CMNs), a probabilistic variant of the mixture-of-experts (MoE) model, are suitable for fast, gradient-free inference and can solve complex classification tasks. CMNs employ linear experts and a softmax gating network. By exploiting conditional conjugacy and P\'olya-Gamma augmentation, we furnish Gaussian likelihoods for the weights of both the linear experts and the gating network. This enables efficient variational updates using coordinate ascent variational inference (CAVI), avoiding traditional gradient-based optimization. We validate this approach by training two-layer CMNs on standard benchmarks from the UCI repository. Our method, CAVI-CMN, achieves competitive and often superior predictive accuracy compared to maximum likelihood estimation (MLE) with backpropagation, while maintaining competitive runtime and full posterior distributions over all model parameters. Moreover, as input size or the number of experts increases, computation time scales competitively with MLE and other gradient-based solutions like black-box variational inference (BBVI), making CAVI-CMN a promising tool for deep, fast, and gradient-free Bayesian networks.

Beyond maximum likelihood estimation (MLE), the standard objective of a language model (LM) that optimizes good examples probabilities, many studies have explored ways that also penalize bad examples for enhancing the quality of output distribution, including unlikelihood training, exponential maximizing average treatment effect (ExMATE), and direct preference optimization (DPO). To systematically compare these methods and further provide a unified recipe for LM optimization, in this paper, we present a unique angle of gradient analysis of loss functions that simultaneously reward good examples and penalize bad ones in LMs. Through both mathematical results and experiments on CausalDialogue and Anthropic HH-RLHF datasets, we identify distinct functional characteristics among these methods. We find that ExMATE serves as a superior surrogate for MLE, and that combining DPO with ExMATE instead of MLE further enhances both the statistical (5-7%) and generative (+18% win rate) performance.

Psychological support hotlines are an effective suicide prevention measure that typically relies on professionals using suicide risk assessment scales to predict individual risk scores. However, the accuracy of scale-based predictive methods for suicide risk assessment can vary widely depending on the expertise of the operator. This limitation underscores the need for more reliable methods, prompting this research's innovative exploration of the use of artificial intelligence to improve the accuracy and efficiency of suicide risk prediction within the context of psychological support hotlines. The study included data from 1,549 subjects from 2015-2017 in China who contacted a psychological support hotline. Each participant was followed for 12 months to identify instances of suicidal behavior. We proposed a novel multi-task learning method that uses the large-scale pre-trained model Whisper for feature extraction and fits psychological scales while predicting the risk of suicide. The proposed method yields a 2.4\% points improvement in F1-score compared to the traditional manual approach based on the psychological scales. Our model demonstrated superior performance compared to the other eight popular models. To our knowledge, this study is the first to apply deep learning to long-term speech data to predict suicide risk in China, indicating grate potential for clinical applications. The source code is publicly available at: \url{https://github.com/songchangwei/Suicide-Risk-Prediction}.

Breast cancer is one of the two cancers responsible for the most deaths in women, with about 42,000 deaths each year in the US. That there are over 300,000 breast cancers newly diagnosed each year suggests that only a fraction of the cancers result in mortality. Thus, most of the women undergo seemingly curative treatment for localized cancers, but a significant later succumb to metastatic disease for which current treatments are only temporizing for the vast majority. The current prognostic metrics are of little actionable value for 4 of the 5 women seemingly cured after local treatment, and many women are exposed to morbid and even mortal adjuvant therapies unnecessarily, with these adjuvant therapies reducing metastatic recurrence by only a third. Thus, there is a need for better prognostics to target aggressive treatment at those who are likely to relapse and spare those who were actually cured. While there is a plethora of molecular and tumor-marker assays in use and under-development to detect recurrence early, these are time consuming, expensive and still often un-validated as to actionable prognostic utility. A different approach would use large data techniques to determine clinical and histopathological parameters that would provide accurate prognostics using existing data. Herein, we report on machine learning, together with grid search and Bayesian Networks to develop algorithms that present a AUC of up to 0.9 in ROC analyses, using only extant data. Such algorithms could be rapidly translated to clinical management as they do not require testing beyond routine tumor evaluations.

Variational continual learning (VCL) is a turn-key learning algorithm that has state-of-the-art performance among the best continual learning models. In our work, we explore an extension of the generalized variational continual learning (GVCL) model, named AutoVCL, which combines task heuristics for informed learning and model optimization. We demonstrate that our model outperforms the standard GVCL with fixed hyperparameters, benefiting from the automatic adjustment of the hyperparameter based on the difficulty and similarity of the incoming task compared to the previous tasks.

What data or environments to use for training to improve downstream performance is a longstanding and very topical question in reinforcement learning. In particular, Unsupervised Environment Design (UED) methods have gained recent attention as their adaptive curricula enable agents to be robust to in- and out-of-distribution tasks. We ask to what extent these methods are themselves robust when applied to a novel setting, closely inspired by a real-world robotics problem. Surprisingly, we find that the state-of-the-art UED methods either do not improve upon the na\"{i}ve baseline of Domain Randomisation (DR), or require substantial hyperparameter tuning to do so. Our analysis shows that this is due to their underlying scoring functions failing to predict intuitive measures of ``learnability'', i.e., in finding the settings that the agent sometimes solves, but not always. Based on this, we instead directly train on levels with high learnability and find that this simple and intuitive approach outperforms UED methods and DR in several binary-outcome environments, including on our domain and the standard UED domain of Minigrid. We further introduce a new adversarial evaluation procedure for directly measuring robustness, closely mirroring the conditional value at risk (CVaR). We open-source all our code and present visualisations of final policies here: https://github.com/amacrutherford/sampling-for-learnability.